Stacked branches and the new world
A few months ago, we rolled out a version of Launchpad that supports and encourages stacked branches. This means medium to large size projects are be able to use Launchpad for hosting their code. I wrote this post shortly before the roll-out, but never got around to blogging it. Hope you enjoy it.
The Problem
As far as code goes, Launchpad thinks in terms of individual Bazaar
branches. Each of these Bazaar branches stands alone, isolated from
other branches. Further, each of these branches has the full history of
whatever project they are a branch of.
That means that every time you push a branch up to Launchpad, you have to push the entire history of your project. That’s not a problem for small projects, but for large projects it’s simply unacceptable. It takes well over an hour for me to push up a branch of Bazaar, for example.
On our side, we simply don’t have the space to store hundreds of copies of the full history of every Free Software project in the world.
The Shape of the Solution
The answer to both of these problems is to share history information
between branches. That way, when you push a new branch, you would only
have to push up the data that’s unique to that branch, the changes that
you’ve made locally. Similarly, we’d only need to store those changes,
rather than another full copy of the history.
Shared Repositories?
Experienced users of Bazaar might be wondering why we didn’t just use
shared repositories. A repository is like a big bucket full of Bazaar
revisions. A repository can be made a shared repository, and branches
can be made to get their revisions from that repository,
rather than from their own. This works fine in many circumstances, but
for a public service like Launchpad, it’s not good enough.
First, people can put poison the bucket. If you can push up a branch, you can write to the shared repository. If you can write to a shared repository at all, then you can write to any revision in that shared repository. This means malicious users could poison any repository that they can push to. If we had a shared repository per project, then malicious users could push up branches that change a project’s history.
A shared repository per person, per project is not so bad, but it’s not so good either. In that model, users would have to push up the full history of a project the first time they contributed a project. Blech. The last thing we want to do is penalize someone for contributing to a new project. Launchpad is supposed to make it easy for you to get started.
Stacked Branches
A stacked branch is a branch that only has a little bit of its own
history. The rest of its history lives in another branch, what we call
the stacked-on branch (I don’t like the name, but all the
better ones are taken). The Bazaar team has been working on making
Bazaar support stacked branches for much of this year.
Stacked branches are exactly what Launchpad needs to get you pushing your branches up quickly. When you push up a stacked branch, you only need write access to the repository of that branch, you only need to read the stacked-on branch. This eliminates the poisoning problem, and allows revisions to be shared safely between branches.
How We Do It
For each project with a development focus (Launchpad’s jargon for an
official trunk), Launchpad has a policy file that tells Bazaar that new
branches should be stacked on that development focus.
So, you make a branch from trunk, do a bit of hacking locally, commit a few revisions and push it up to Launchpad as your own branch, say lp:~jml/tribunal/awesome-new-feature. When you push, Bazaar asks Launchpad where it should stack the branch, and Launchpad says lp:~jml/tribunal/trunk. Bazaar then pushes only the new revisions I’ve added to my “awesome-new-feature” branch.
We do a lot of stuff behind the scenes to make sure the branches work over HTTP, bzr+ssh, SFTP and to make sure that mirrored branches work just as well as hosted branches and so on. I hope that most of this is invisible to you, and that you simply notice that pushing big branches to Launchpad suddenly got a lot faster.
The Future
By making it possible to quickly push big branches to Launchpad, we’ve
entered a new era. Projects like Bazaar and Launchpad itself can start
actually hosting their branches on Launchpad. For other projects (like
Twisted, Python or even GNOME), hosting branches on Launchpad is now an
option that can be seriously considered.
For my part, stacked branches are already changing the way I work. I’m now using the product that I helped build each and every day as part of my job. Further, my colleagues and my boss are also using it. This quickly exposes rough edges and raises obvious ideas for improvement in the rest of the Code site. It also helps me see which bugs are the really important ones.
More importantly, it’s reminded me that Launchpad actually is pretty cool. I’m using it more, and looking forward to the day when I can really, genuinely say that Launchpad is a joy to use.